1DevTool field report — 7 issues from heavy multi-agent use (with measurements & minimal proposals)

Hi anh Khoa,

Environment: macOS 15, CLI 1devtool-agent-v7, testing on a large repository (8,292 files).
7 Core Issues & Proposed Fixes
1. Unlimited Terminal Tab Accumulation (Tab Leaks)
⚬ Issue: 30 tabs were left hanging after one session. The main app has the capability to close tabs, but the CLI and HTTP bridge do not expose it (POST returns 404). The stop command currently only works for team-associated tabs.
⚬ Proposal: Update the stop command to accept --terminal=, or add a new terminal close --id= command.
2. Permanent Hang at "Completion needs confirmation"
⚬ Issue: When an agent finishes and parks in this state, all subsequent dispatches hang at "delivering" forever. A team can only survive exactly one round of work.
⚬ Proposal:
⚬ Document the undocumented resolve command in --help (which actually fixes this).
⚬ Auto-confirm when the agent writes its result file.
⚬ Allow clearing leftover attention badges after an orchestration is closed.
3. Link vs. Headless Deadlock
⚬ Issue: Running headless is refused if an active link exists. However, agent-created links default to requiring confirmation, causing messages to sit permanently in the queued state. The error returns exitCode:None, mimicking a broken worker and costing hours of debugging.
⚬ Proposal: If headless mode is blocked by a link, enforce that the link must be in a flowing mode. Alternatively, allow headless with a warning, or explicitly state "link is awaiting approval" in the error.
4. Silent Git Worktree Creation & Timeout
⚬ Issue: Using category=implement|debug triggers a silent git worktree creation. For a large repo, this takes >13.7s, triggering an internal timeout and failing with a vague "Can't run" error.
⚬ Proposal: Document which categories create worktrees, scale the timeout limit based on repo size, and surface the actual timeout error rather than "Can't run".
5. Headless Codex Crashes in 96ms (Missing node PATH)
⚬ Issue: The app spawns a shim but its PATH lacks node (despite it being installed on the system). It fails with "env: node: No such file" and dies instantly.
⚬ Proposal: Inherit the user's login PATH when spawning processes, or resolve node absolutely. Report "node not found on PATH" instead of raw environment errors.
6. False Positives from terminal submit
⚬ Issue: The command returns rc=0 regardless of whether the prompt runs immediately or queues for 10+ minutes. External orchestrators cannot distinguish between "delivered" and "running" because terminal list --json lacks a busy/idle field.
⚬ Proposal: Make submit return a receipt (submissionId + tab state) or expose a read path for external callers to verify tab status.
7. External "Machine Terminals" (e.g., Ghostty) Lack Attribution
⚬ Issue: Terminals opened externally do not appear in terminal list and lack a ONEDEVTOOL_TERMINAL_ID. Consequently, resolve commands fail due to unrecognized ownership.
⚬ Proposal: Extend the attach protocol so external terminals can opt-in to attribution, or explicitly document that machine terminals cannot make attributed calls.
Additional Polish & Positives
⚬ Polish Item: Permission flags are highly fragmented across CLIs (e.g., --dangerously-skip-permissions vs. --always-approve). Using the wrong one causes crashes or silent blocks. Proposal: Create a single, unified intent flag (like --auto-approve) that 1DevTool translates internally for each specific CLI.
⚬ What Works Well:
⚬ The socket-based terminal list / view / submit path is excellent for external orchestration.
⚬ Team orchestration with per-member transcripts is highly practical.
⚬ With issues (1), (2), and (6) resolved, 1DevTool would be fully ready for end-to-end automation. Available to provide journals, run records, or reproduce bugs on a call if needed.

Happy to provide journals, run records, or reproduce any of the above on a call.

Thanks,
Phương Nam